2.1 Why data collection comes first

Every later section of the course takes a table of counts and extracts from it a measure of frequency or association, with a confidence interval and perhaps an adjustment for confounding. All of that machinery assumes that the counts mean what they say: that a person recorded as a case had the disease, that a person recorded as exposed was exposed, and that the people counted are the people the study meant to count. No analysis can repair data that fail these assumptions. A confidence interval measures the random error of a well-collected sample; it says nothing about systematic error, and a large study with biased measurement gives a narrow interval around the wrong value.

This is the sense in which data collection has primacy. The decisions that most affect the validity of an epidemiological study are made before any analysis: who is eligible, how cases are defined and found, how exposure is measured, how many people are lost or refuse. The statistician’s part in them is to say, in numbers, what each choice will cost.

The protocol

A study begins with a written protocol, fixed before data collection, which states:

1.

the question, as a comparison of specified measures in a specified population (for instance, the rate ratio of hip fracture for current users of a drug against non-users, among women aged 65 and over);

2.

the design and the source population;

3.

the case definition and how cases will be ascertained;

4.

the exposure definition, its time window and how it will be measured;

5.

the variables to be recorded as possible confounders;

6.

the sample size, with the calculation behind it;

7.

the planned analysis, including what will be done about missing data.

Writing the analysis down in advance matters. With many possible definitions of exposure, cut-points and adjustment sets, an analysis chosen after looking at the data can find an association that is an artefact of the choice.

Sources of data

Epidemiological data come either from routine sources, collected for other purposes (death certificates and vital registration, disease registries such as cancer registries, notifications of infectious disease, hospital discharge records, insurance and prescription databases, censuses) or from data collected for the study (interviews, self-completed questionnaires, physical examination, laboratory measurement of biological samples). Routine data are cheap and cover whole populations, but their definitions were set by someone else and their quality varies with place and time: a rise in the recorded rate of a disease may reflect a new diagnostic test or a change in coding rules. Study data can be made to fit the question, at a cost that limits the number of people measured.

A pilot study on a small sample tests the questionnaire, the field procedures and the timing before the main study starts, and gives the first estimates of the quantities (prevalence of exposure, response rate, variance) that the sample size calculation needs.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.